PHYSICAL REVIEW E 80, 046201 2009

Rotating bouncing disks, tossing pizza dough, and the behavior of ultrasonic motors
Kuang-Chen Liu, James Friend, and Leslie Yeo
MicroNanophysics Research Laboratory, Department of Mechanical Engineering, Monash University, Victoria 3800, Australia
Received 4 May 2009; revised manuscript received 1 August 2009; published 1 October 2009
Pizza tossing and certain forms of standing-wave ultrasonic motors SWUMs share a similar process for
converting reciprocating input into continuous rotary motion. We show that the key features of this motion
conversion process such as collision, separation and friction coupling are captured by the dynamics of a disk
bouncing on a vibrating platform. The model shows that the linear or helical hand motions commonly used by
pizza chefs and dough-toss performers for single tosses maximize energy efficiency and the doughs airborne
rotational speed; on the other hand, the semielliptical hand motions used for multiple tosses make it easier to
maintain dough rotation at the maximum speed. The systems bifurcation diagram and basins of attraction also
provide a physical basis for understanding the peculiar behavior of SWUMs and provide a means to design
them. The model is able to explain the apparently chaotic oscillations that occur in SWUMs and predict the
observed trends in steady-state speed and stall torque as preload is increased.
DOI: 10.1103/PhysRevE.80.046201

PACS numbers: 05.45.a, 45.20.dc, 46.40.f, 47.20.Ky

I. INTRODUCTION

Like many nonlinear dynamical systems, the seemingly

simple equations that govern a ball bouncing on a vibrating
platform under a constant gravitational field describe very
complex behavior. A variety of interesting phenomena accompany the bouncing ball system, including noise-sensitive
hysteresis loops 1, period doubling route to chaos, and
eventually periodic orbits known as the sticking solution,
or complete chattering 2,3. Due to its physical simplicity
and rich nonlinear behavior, the system has found a variety
of applications, from the dynamics of two-dimensional
granular gases 4, high energy ball milling 5, to its use as
a pedagogical demonstration of chaos 6.
Regardless of the various complex bouncing dynamics
that it displays, the response of the traditional bouncing ball
system is ultimately an oscillatory one: the mean vertical
velocity is zero, and, as long as the collision is not perfectly
elastic, the displacement of the ball is bounded. However,
through a simple extension of the system by adding a rotational degree of freedom and an angular component to the
platform vibration, a different set of phenomena is made possible: stick-slip rotation while the disk and platform are in
contact, impulsive frictional torque imparted at each collision, and the potential for nonzero mean angular velocity and
unbounded rotary motion.
This process of motion transferthe conversion of reciprocating motion into continuous rotary motion in the modified systemis intimately related to pizza tossing and the
stator-rotor interaction in a class of standing-wave ultrasonic
motors SWUMs 79. As illustrated in Fig. 1, the system
may be represented as a disk or a pizza dough with angular
and linear x displacements, bouncing on an oscillating platform or a pair of dough-tossing hands with a combined angular and linear oscillation s described by

t = sint +

and

st = A sint.

Figure 2 shows a family of potential trajectories traced by a

dough-tossing hand or a SWUM stator as the amplitude ratio
1539-3755/2009/804/04620111

L = ae / A = tan L and phase lag are varied ae is the effective friction contact radius.
Of the many different hand trajectories that may be used
to toss a pizza, we observe that distinct hand motions are
used in two different dough-tossing modes. In the first mode
10, the dough begins each cycle at rest relative to the hand:
the dough is launched, caught upon its descent, allowed to
come to rest, and the process is repeated. Since each toss
begins with the same initial conditions, the process can be
considered as a chain of single tosses. In this case, we observe that a linear trajectory resembling Fig. 2c is employed. In the second mode 11, the dough is not allowed to
come to rest after each collision, and thus the rotation of the
dough is maintained over multiple tosses. In this case, the
tossing motion traces a semielliptical trajectory resembling
Fig. 2b.
One of our key goals in this paper is to investigate why
the particular hand motions are adopted by dough-tossing
performers for the two dough-tossing modes. Do these hand

bouncing disk/pizza dough

datum

platform/hand trajectory

FIG. 1. Color online The displacement variables of the

bouncing-disk system and pizza tossing. The trajectories traced by a
point on the oscillating platform and a dough-tossing hand are given
by Eqs. 1, which describe a family of closed curves on the surface
of a cylinder.

046201-1

2009 The American Physical Society

PHYSICAL REVIEW E 80, 046201 2009

LIU, FRIEND, AND YEO

a

b

 0.5  0.25

c

d

0

e

f 

L  0.05 L  0.25 L  0.45

FIG. 2. Color online Potential dough-tossing motions and

SWUM stator trajectories: ac the effect of varying phase lag
with L = / 4, and df the effect of varying amplitude ratio L
= tanL with = / 4.

motions provide any advantages in terms of the effort required, the rotary speed reached by the dough, and ease of
handling? Noting that the operation of SWUMs can be seen
as a continuous sequence of multiple pizza tosses, answers to
the above questions about pizza tossingoriginally conceived as merely a pedagogical toolwill also help us better
understand the underlying motion transfer process in
SWUMs: the generation of continuous rotation from an oscillatory input.
A number of attempts have been made to model SWUMs
over the years 7,12,13, however, these models have so far
assumed that the vertical motion of the rotor is negligible.
This assumption may be valid when the rotor preload dominates over the inertial force of the stator acceleration and the
rotor operates in the nonbouncing regime. More precisely, in
terms of the dimensionless forcing parameter = A2 / g,
where g is the rotor acceleration due to preload, the rotor will
remain in contact with the stator when 1 / . However,
most motors operate with 1 / . For example, our SWUM
8 operating at 70 kHz with the rotor under gravitational
acceleration has a lower estimate of 6 if we assume a
stator vibration of merely 1 nm; Tsujinos SWUM 14 with
a stator vibration of 5 m at 55 kHz and a rotor under a
preload acceleration of 9 103 m / s2 has 70. As will be
shown, the bouncing-disk model reproduces key qualitative
features of the motors dynamics, including the seemingly
chaotic oscillations of the rotors transient speed curves and
the effect of preload on steady-state speed and stall torque.
The structure of the current paper is as follows: in Sec. II
we describe details of our bouncing-disk model; in Sec. III
we use our model to investigate the optimal hand motion for
single tosses; in Sec. IV we investigate the best way to maintain dough rotation over multiple tosses and its implication
for SWUMs by considering how various bouncing-disk orbits e.g., periodic, chaotic, or chattering and their basins of
attraction affect the motion transfer process; and in Sec. V,
we compare our models predictions with results from a prototype motor, showing that the bouncing-disk model is able
to account for important SWUM characteristics that could
not be explained by existing models.
Note that we neglect dough deformation in this paper for
the following reasons: 1 our focus is on pizza tossing as a
method for imparting rotary motion rather than dough shaping, 2 the typical rate of dough deformation is low 3%
increase in diameter per toss, see Appendix, and 3 pizza
tossing is used as a pedagogical tool for understanding the
rotor dynamics in SWUMs. Dough plasticity only has a minor effect on the dynamics and will not be considered further

here. Although we assume that the rotor of SWUM is under

constant preload, our model would still be valid for sprung
rotors if the spring with stiffness k operates purely in compression and the springs mean displacement l0 is much
greater than the rotors peak-to-peak displacement amplitude. Under such conditions, the spring force is effectively
constant and the rotors acceleration due to the spring preload is given by g = kl0 / m.
In a previous publication 15, we have briefly reported
the application of the bouncing-disk model to pizza tossing,
and its potential implications for SWUMs. In this paper, we
provide the full details for the modeling of single and multiple pizza tosses, and we make an in-depth comparison between our theoretical predictions and the experimental results of SWUMs.
II. THE BOUNCING-DISK MODEL
A. Equations of motion

As an extension of the traditional bouncing ball problem,

the vertical component of the bouncing-disk system has the
same assumptions and governing equations as previous work
3: collisions have zero duration and the coefficient of restitution is independent of the impact velocity. For the rotary
component, we assume that the contact pressure is uniform
and that the frictional torque can be modeled using Coulomb
friction with a constant coefficient of friction acting at an
effective contact radius ae we assume that the kinetic and
static coefficients of friction are equal. The resulting governing equations for the motion of the disk are

t t ,
ma2g = T f + H
n
n

2a

n=1

and

mx = mg + N + Fnt tn,

2b

n=1

where N is the normal contact force, T f is the frictional

are the linear and angular impulse at time
torque, Fn and H
n
tn from the disks nth collision with the platform, is the
Dirac delta function, m is the disks mass, and ag is the
radius of gyration. The platform displacements s , prescribed by Eqs. 1 enters Eqs. 2 through the following
; their functional forms are
forcing terms: N, T f , Fn, and H
n
described below.
During the contact phase, the normal contact force N assumes a value such that x = s. However, in the absence of
adhesion forces N must remain non-negative; thus the rotor
and the stator separates if N falls to zero when s g. The
functional form of N is then
N=

0,

s g or x s

mg + s, s g and x = s

Depending on the normal contact force N, the relative

angular velocity and the angular acceleration of the
stator, the torque acting on the rotor can arise from either

046201-2

PHYSICAL REVIEW E 80, 046201 2009

ROTATING BOUNCING DISKS, TOSSING PIZZA DOUGH,

static or kinetic friction. When the sliding speed is zero

= 0 and the angular acceleration is within or entering
the static friction envelope Senv = : env where env
= aeN / ma2g, the frictional torque is static and it exactly
balances the rotors inertia so that = . If the sliding speed
is nonzero 0, or the angular acceleration is outside
or exiting the static friction envelope, then the frictional
torque is kinetic and it acts to oppose the sliding motion. The
above description is partially summarized by the following
equation:

Tf =

= and env
= and env
= and env

ma2g ,

aeN,
aeN,

aeN sgn ,

parabolic
flight

Separation

t+n

tn

Ftdt = mx+ x,

False rebound
True contact
speed is
force falls False
zero
below
C A
sliding
zero
True
contact
sticking
False
True contact
frictional
relative
torque is & angular
Attachment
below the
speed is
static limit
zero
S

4
=
H
n

t+n

tn

ae sgn Ftdt = ae sgn Fn .

Note that the frictional torque may only be present over a

fraction of the collision, vanishing when the relative angular
velocity between the rotor and the stator is reduced to zero
static friction is ignored since the collision is short. If a cap
, scenarios where the relative rotation of
is not imposed on H
n
the rotor reverses direction after impact would occur. The
maximum transferable impulse is thus limited to the amount
that would result in + = 0, which is

and the frictional angular impulse is given by

10

Except for gravity, the forces experienced by the bouncing

disk are all associated with some form of discontinuity: the
cause sharp changes to the disks
impulsive forces Fn , H
n
velocity, the contact force can only support compressional
loads N 0, and the frictional torque can switch between
static and kinetic friction and has a jump discontinuity when
relative sliding velocity changes sign aeN sgn .
The result of these discontinuities is that the disk described
by Eq. 2 experiences four distinct phases: 1 parabolic
flight, 2 impact, 3 sticking contact, and 4 sliding contact, each of which may be solved analytically. The following paragraph outlines how the four phases are sequenced as
a solution to Eq. 2 and the whole process is summarized in
Fig. 3.
During parabolic flight, gravity is the only acting force so
that xt = xn + vnt tn 2g t tn2 and t = n. The phase
ends when the disk collides with the platform, where the
time of collision is determined by solving for tn+1 such that
xtn+1 = stn+1

= ma2 ,
H
n,p
g

H
, if H
H
n,p
n,a
n,p

B. Method of solution

requires
The determination of the frictional impulse H
n
and the maxiconsideration of the available impulse H
n,a
is
. The available impulse H
mum possible impulse H
n,p
n,a
the torsional impulse that would be transmitted if sliding
friction were present over the whole collision
=
H
n,a

H
, if H
H
n,a
n,p
n,a

Substituting Eq. 6 into Eq. 5, we obtain the expression for

the nth axial impulse
Fn = mx+ x = m1 + s x.

impact

FIG. 3. Color online Model simulation flow chart. The inset

a shows a potential trajectory of the bouncing disk dashed line
and the oscillating platform solid line for one cycle of simulation,
starting with collision c, attachment a, and ending with separation s.

where the superscripts , + are used to denote, respectively,

a quantity before and after the collision. Since we are modeling the collision with a speed-independent coefficient of
restitution , the relationship between the preimpact and
postimpact relative vertical velocity is
s x+ = s x.

Collision

(a)

where sgn is the signum function.

The axial impulse of the nth collision Fn is given by
Fn =

g
xn + vntn+1 tn tn+1 tn2 = A sintn+1.
2

11

After determining the postimpact velocities through Eqs. 6

and 10, if the rebound speed x s is zero and the contact
force is positive N 0, the system enters one of two contact
phases, otherwise the disk is relaunched into parabolic flight.
The two contact phases differ in terms of whether the friction
is static or kinetic. If the disk satisfies the condition described in Eq. 4 for static friction, then = ; otherwise, the
system undergoes sliding contact and the angular rotation of
the disk is determined by solving

046201-3

PHYSICAL REVIEW E 80, 046201 2009

LIU, FRIEND, AND YEO

TABLE I. Full specification of a single toss

rad

L
rad

c
m

rad/s

ae
m

ag
m

g
m/s

0.55 0.14

0.25 0.05

15 5

0.135 0.015

0.11 0.01

0.6 0.1

9.8

ma2g = aeN sgn .

and the coefficient of restitution are not part of the specification because we are only concerned with a single launch
cycle and the dough always begins each toss at rest with t0
= 3T / 4 = 3 / 2 i.e., the lowest point of the tossing motion.
The parameter values we use in this investigation are shown
in Table I. The coefficient of friction between skin and
dough is estimated to be = 0.6 0.1 through the inclined
plane test a baked bread begins to slip on a lightly floured
hand at a slope of 30 5; the gravitational acceleration
g = 9.8 m / s2. The rest of the parameters are based on a video
of pizza tossing recorded at a local pizza shop 10: , ae, ag,
and c are estimated by visual inspection; L and , however,
are determined from the estimated vertical and angular
launch speeds: xsep = 3.1 0.3 m / s and sep = 14 1.5 rad/ s.
Assuming that there is no slip between the dough and the
hands,

12

The two contact phases switch between each other according

to the following conditions: stick changes to slip if the platforms angular acceleration exceeds the static friction limit,
and slip changes to stick if the sliding speed reaches zero.
Both contact phases end when N falls to zero and the disk is
launched into parabolic flight.
III. SINGLE TOSSES OF PIZZA DOUGH
A. Performance measures and parameters

We simulate our bouncing-disk model for a single launch

cycle to investigate what the best hand motion is for single
tosses as and L are varied. To compare the different tossing
motions, the following performance measures are used: 1
the airborne rotational speed f reached by the pizza dough,
2 the energy efficiency or ratio of the rotational kinetic
energy gained by the dough over the total energy input due to
the tossing motion, and 3 the speed ratio of the doughs
airborne angular velocity f over the maximum rotary tossing speed .
Noting that the amplitude ratio L = tan L = ae / A, and
that the arc length of the linear dough-tossing trajectory c
= A2 + ae2, we can express A and as functions of L
and c
ae,A = csin L,cos L.

xsep = c cos L

and L are thus given by

2 =

5
6

10

7
8

20
0

L
rad

3
8

0.4

n0

0.3

1
7

0.2

0
0

2
3

0.1

3
8

L
rad

14

ae sep
.
xsep

The effects of varying the amplitude ratio L and the

phase lag are shown in Figs. 4a4c. For clarity, we only
show the results for 0 since f , , and all possess
some form of translation symmetry: for f and , f + =
f, and for , f + = f. We have also limited the
domain of L to 0 , cos1g / c2 rather than the full range
of 0 , / 2 because the vertical force exerted by the tossing

d
0.00

5
6

0.5
1.
0

L
rad

0.2

e
0.00
25

rotary speed  f
rads

B. Results and discussion

phase sep2

sepae
c

L = tan1

c
0.00
1. n  01
2
0.5
3

speed ratio

n  0 12
3

xsep
c

13

b
0.00
0.5
energy efficiency

rotary speed  f
rads

10

and

Specifying A , in terms of c , L has the following advantages: by keeping c constant, the stroke length of the
dough-tossing motion remains bounded as we vary the amplitude ratio L, and by varying L between 0 and / 2 all
possible L from 0 to are explored.
The full specification of a single toss in our bouncing-disk
model can be determined by the following set of independent
parameters: , L, c, , ae, ag, , and g. Initial conditions

a
0.00
20

sep = c sin L/ae ,

and

L
cos1
gc2

0.1

3
8

0
0

L
rad

3
8

20
15
increasing

10
5
0
0

L
rad

3
8

FIG. 4. Color online The performance of different pizza tossing techniquesas measured by a dough rotation speed f , b energy
efficiency , and c dough-to-hand speed ratio when the amplitude ratio L is varied at a phase lag of = n / 8. d f is the dough
rotation speed at the phase of separation sep, which is a function of L with a range of sin1g / c2 , / 4. e The effect of on f ; the
maximum f occur at greater L as is increased 0.5, 0.6 and 0.7 for the three curves.
046201-4

PHYSICAL REVIEW E 80, 046201 2009

ROTATING BOUNCING DISKS, TOSSING PIZZA DOUGH,

a

s, x
ms

TABLE II. Effects of input parameter uncertainties

ae

ag

14

0.01

0.01

motion is insufficient for overcoming gravity when L

cos1g / c2, which means the dough never leaves the
hand.
Though our model is simple, the essential physics of the
torque transfer process in pizza tossing are included. The
predicted dough rotation speed is 14 5 rad/ s for the single
toss specified by Table I, which is in good agreement with
the average airborne angular speed of 14 1.5 rad/ s in the
first four pizza tosses of the video footage 10. The effects
of the uncertainties in the input parameters to our model are
shown in Table II.
From Figs. 4a4c, we can see that the optimal parameters are = 0 or , and L = 0.323 L = 1.6, where the
maximum airborne speed f = 22 rad/ s and energy efficiency = 0.42 are reached; the speed ratio at these points
= 0.97 have a negligible difference from max = 0.99 at
L , = 0.132 , 0. The optimal hand motions predicted by
our bouncing-disk model are linear trajectories with = 0 as
shown in Fig. 2c, which are precisely the motions we observe in actual pizza tossing. We note, however, that there is
a significant difference between the optimal L = 0.323 predicted by our model and the observed L
= 0.175 0.045. We will first explain why L ,
= 0.323 , 0 is optimal in our model before returning to
discuss this discrepancy.
When the contact force and the coefficient of friction are
sufficiently large, the airborne rotation speed of the dough
will be the same as the rotation speed of the tossing motion
at the point of separation when s falls below g. In terms of
the nondimensional phase = t, the point of maximum hand
rotation speed max and the point of separation sep are,
respectively, given by

max = + n,
and

where n = 0,1. . .

sep = sin11/.

15

The amplitude ratio tan L = ae / A affects three things:

the point of separation sep see Fig. 4d, the maximum
hand rotation speed , and the normal contact force N
= mg + x; the phase lag, on the other hand, controls the point
at which the maximum hand rotation speed occurs max . If
we let L / 2, we increase the hand rotation speed ,
however the dough will be sliding at sep due to reduced
normal contact force; if we let L 0, there will be little
rotary hand motion compare the three columns in Fig. 5.
The optimal amplitude ratio L = 0.323 thus results from
the compromise between the goal of minimizing slip and the
goal of maximizing . The optimal phase lag = 0, on the
other hand, results from the balance between the torque
transmitted at the beginning 0 and the end sep of the


, 
rads

Uncertainties rad/s due to

f
rad/s

b

sep

c

sep
0.1
0.4
0.2

0.6
0

0.6
0

0.6
0

0.2

0.4

0.1
0 75 1 1 25 1 5 0 75 1 1 25
1 1 5 0 75 1 1 25 1 5
1
20
20
20
1
2
2
2

0.6
0

0.6
0

0.6
0

20

sep

20

20

0.75 1. 1.25 1.5 0.75 1. 1.25 1.5 0.75 1. 1.25 1.5

phase 2

FIG. 5. Color online Details of various hand motions dashed

curves and the resulting dough trajectories solid curves. Two trajectories with different phase lags 1 = 0 and 2 = / 2 are shown
for each of the following amplitude ratios L: a 0.132, b
0.323, and c 0.426. The effect of choosing a suboptimal L,
such as slip at high L, and low hand rotation rate at low L,
can be seen by comparing the trajectories in a and c with the
optimal trajectory with L = 0.323 in b.

contact phase. If 0, max is shifted away from sep to the

left and thus the torque is reduced as we approach separation;
if 0, max is shifted toward sep on the right but there
would initially be reverse torque at 0.
Our model suggests that the linear hand motions observed
in the single toss video = 0 are chosen to maximize dough
rotation and energy efficiency. However, the observed amplitude ratio L = 0.175 0.045 is much lower than the optimal value of L = 0.323 predicted by our model, which
may be caused by other factors that influence the choice of
the amplitude ratio L. For example, a dough-toss performer
may be aiming for a particular toss height which requires a
smaller L see Fig. 5 for the effect of L on toss height.
Furthermore, Fig. 4e shows that the optimal L shifts toward zero as decreases. Since the coefficient of friction
depends on the amount of dry flour present on the handsa
factor difficult to controla dough toss performer may thus
choose a low L to minimize slip.
IV. MULTIPLE TOSSES OF PIZZA DOUGH AND SWUM
BEHAVIOR

In this section we investigate the underlying motion transfer process that applies to both SWUMs and multiple pizza
tosses. We will thus refer to both the pizza dough and the
SWUM rotor as the disk, and both the dough-tossing hand
and the SWUM stator as the platform.
Most of the angular momentum transfer between the platform and the disk occurs during impact. Much information
about the motion transfer process can thus be gained by
studying the next collision map of the bouncing-disk system: the location of attracting orbits will affect the steadystate rotation speed, and the size of their basins of attractions
will affect the sensitivity of SWUMs to perturbations and the
ease at which multiple pizza tosses may be executed. The
following state variables are used in our next collision map:
the relative axial collision velocity wA = x s, the relative angular collision velocity wT = , and the phase at impact
imp = timp. The nondimensionalized form of the map has
four parameters: , , as previously defined, and an angular

046201-5

PHYSICAL REVIEW E 80, 046201 2009

a

bc d e
b

imp2

w AA

LIU, FRIEND, AND YEO

a
0.4
0.2
0
0.2
0.4
0.35
a

bc

de

during the chattering phase and the process is repeated. On

the other hand, the periodic orbits better match the dynamics
of the multiple pizza tosses, and will thus be the focus of our
discussion.
acts always to
Since the frictional angular impulse H
n
reduce the relative angular impact speed wT between the
disk and the platform, wT will eventually be zero for any
period 1 orbit. That is,

0.

0.5
1.

1.5
0.45 0.55
0.35 0.45
Axial forcing parameter  
bc

0.55

d e
d



1.



a
bc d e
1.
0.5
0.5
0.
0.0
0.5
0.
1.
0.35 0.45 0.55
0.35 0.45 0.55
Axial forcing parameter  

c

p1,ss = timp = cosimp + ,

FIG. 6. Color online Bifurcation diagrams = 0.5 showing

period doubling routes to chaos dark gray and the evolution of the
chattering mode light gray as is increased for: a the phase of
impact imp, b the relative vertical collision speed, cd the
normalized angular collision speed when c , T = 0 , 100 and
d , T = 0.26 , 1. Note that L is not fixed. The basins of
attraction at ae are shown in Fig. 7. The axial forcing parameter
n at the first five bifurcation points are 0.486, 0.531, 0.5402,
nn1
0.5421, and 0.5425. The ratio n+1n of the distance between the
first few bifurcation points are 4.891, 4.842, and 4.750, which approaches the Feigenbaum constant 4.669

forcing parameter T = a2g2 / aeg that can be interpreted

as the ratio of the inertial torque to the frictional torque due
to the rotor preload. In the next two sections we will discuss
the implications of the systems bifurcation plots and basins
of attraction on pizza tossing and SWUMs.
A. Bifurcation diagram

The use of Coulomb friction in our bouncing-disk model

means that the angular component of the disk is affected by
its vertical motion but not vice versa. The governing equation for the vertical component of our bouncing-disk system
is thus identical to the traditional bouncing ball system. We
plot the bifurcation diagram for = 0.5 by following the
period 1 attractor 16 at = 0.335, which undergoes a period
doubling route to chaos as is increased. The resulting bifurcation diagram for imp, shown in Fig. 6a, is consistent
with Tufillaros 2.
Also shown in Fig. 6 is the evolution of the chattering
orbit, which coexists with the periodic orbit as a separate
attractor in the phase space. A key feature of the chattering
orbit is the occurrence of rapid collisions with decaying impact velocity and flight duration. In the idealized bouncingdisk system, it is possible for the disk to come to rest on the
platform after completing an infinite number of collisions in
finite time; to avoid this computational supertask 17, we
artificially set wA to zero if the impact speed is small
wA / A 106 or if there are 20 consecutive short duration flights / 2 104. Although the complexity of the
chattering orbits grows as is increased, they still appear to
be eventually periodic. For 0.52, the chattering orbits
follow a pattern similar to those of single pizza tosses: following a long toss, the dough comes to rest on the hand

16

where p1,ss is a period 1 steady-state dough rotational velocity, and timp is the platform rotation velocity at impact.
The maximum steady-state rotational speed p1,ss = will
be achieved if = imp. In Fig. 6a, we can see that period
1 orbits occur for 1 / 3 0.486, and the phase of impact
imp / 2 starts at 0 and shifts to 0.130 at the first point of
bifurcation, thus the stator motion needed to achieve the
maximum rotation velocity can vary from a linear to a
semielliptical trajectory depending on .
For example, a linear trajectory with = 0 would give the
maximum rotation speed p1,ss = when = 1 / 3 and decrease to p1,ss = cos0.26 when = 0.486, as shown in
the bifurcation diagram of / for , T = 0 , 100 in
Fig. 6c. In contrast, a semielliptical trajectory with =
0.26 gives a low rotary speed when = 1 / 3 and reaches
the maximum speed when = 0.486, as shown in Fig. 6d.
In the period 2 region 0.486 0.531, Eq. 16 no
longer applies as the steady-state orbit alternates between
two collision states: the large-wA branch with a decreasing
imp, and the small-wA branch with an increasing imp see
Fig. 6b. Since timp of the large-wA branch has a greater
influence on the disks steady-state rotational speed, as imp
shifts toward zero on the large-wA branch, the average rotational speed rises for Fig. 6c and falls for Fig. 6d until
reaching the local extremum corresponding to the period 4
bifurcation.
As the disk visits an increasing number of points in the
phase space post period 4 bifurcation, and into the chaotic
and chattering regime, the rotational dynamics of the disk is
increasingly affected by T. For a large T, the rotary inertia
of the disk dominates over the frictional torque, and the disk
rotates at an average speed with small perturbations due to
transferred at different times . For a
angular impulses H
n
imp
small T, however, the frictional torque dominates and the
rotary speed of the disk oscillates wildly as the disk tracks
timp more closely. The effect of T on this oscillation of
the speed can be seen by comparing Fig. 6c where T
= 100 and Fig. 6d where T = 1. Note that the bifurcation
diagram may appear to have three branches in what should
be a period 4 region of Fig. 6d. However, this is caused by
the one-dimensional projection of orbits that exist in threedimensional space imp , wA , wT.
Past SWUM researchers 7,13 have assumed that the rotor can be considered stationary relative to the stator, leading
to the conclusion that contact always occurs as the stator
reaches its maximum vertical displacement / 2, and
that an elliptical trajectorysuch as shown in Fig. 2a, with

046201-6

PHYSICAL REVIEW E 80, 046201 2009

collision velocity
w A

A

ROTATING BOUNCING DISKS, TOSSING PIZZA DOUGH,

a
0
2.5
5.
0.5





b
0



 

2.5
5.
0.5 0.5

c
0


2.5





d
0
chattering
orbit

2.5

    
 

5.
5.
0.5 0.5
0
0.5 0.5
phase of collision 
2

chaotic
orbit

e
0

2.5
5.
0.5 0.5

  
  
  

chattering orbit

0.5

FIG. 7. Color online The basins of attraction for the following values of : a 0.34, b 0.486, c 0.50, d 0.55, and e 0.60. The white
regions are the basins of attraction of the chattering orbits marked by black s in ac, red dots in d, and black dots in e. The black
regions are the basins of attraction of the attractors that undergo the period doubling route to chaos marked by white in ac and
yellow dots in d.

= max = / 2is the optimal stator motion. We have

shown, however, that when the vertical rotor motion is taken
into account, the maximum steady-state rotational speed is
achieved by choosing the right phase lag = imp for a
given in the period 1 region 0.333 0.486. Contrary
to the assumption of SWUM researchers, neither imp is necessarily / 2 nor is elliptical motion the best stator trajectory.
B. Basins of attraction

We can see from the previous section that in order to

achieve the maximum dough rotation speed, one should
choose a hand motion within the period 1 regime 0.333
0.486 for multiple pizza tosses. We can further compare the relative ease at which a hand motion maintains
dough rotation over multiple tosses by examining the effect
of on the basins of attraction; a larger basin of attraction
implies that the orbit is more tolerant of variability in the
hand motion. In Fig. 7, the basin of attraction of the period 1
orbit at = 0.34 only covers 11.6% of the phase space
shown; as is increased, the basin of attraction widens to
39.1% of the phase space shown at = 0.486. We can thus
see that the use of semielliptical hand motion = 0.26
with 0.486 has the advantage of being easier for maintaining the period 1 orbit required for maximum dough rotation speed.
The wide basins of attraction of the chattering orbits make
them far easier to perform than period 1 multiple tosses. The
fact that the wA = 0 axis is deep within the basins of chattering orbits shows that period 1 orbits cannot be attained for a
dough initially at rest if the tossing motion purely follows
Eq. 1. This explains why dough-toss performers do not
start multiple tosses with semielliptical hand motion right
from the beginning, but instead employ a linear hand motion
for the first few tosses.
Although it is unlikely that dough-toss performers will
need to consult Fig. 6 and 7 to avoid chaotic and chattering
orbits, the bifurcation diagrams and the basins of attraction
provides SWUM researchers useful insights into the physical
behavior of SWUMs. For example, SWUMs should be designed to operate away from the chaotic regime e.g.,
= 0.55 in Fig. 7d because the extended region of the phase
space visited by the disk reduces the rotor speed, and the
fractal basin boundary of the attractors makes the motor behavior unpredictable. The maximum rotor speed can be
achieved in the period 1 regime, though complex control
may be needed to actively adjust the stator motion to accommodate the evolving stator-rotor dynamics, since the basin of

attraction for period 1 orbits do not cover SWUMs usual

initial condition. For 0.561, we see a sudden widening of
the phase space explored by the orbits; the strange attractor
of the chaotic orbits merges with the chattering orbit and the
basin of attraction appears to fill the entire phase space see
Fig. 6d. Although these orbits are eventually periodic, they
differ from the low period chattering orbits in Figs.
7a7c: they have a longer period, and they follow the
shadow of the chaotic strange attractor. SWUMs operated in
this regime will not reach the maximum rotor speed, however, the wide basin of attraction suggests that they will operate most consistently. In Sec. V, our simulation results suggests that our prototype SWUM predominantly operates in
this regime.

V. DYNAMICS OF SWUMS

The rotor and stator of our SWUM are, respectively, a

20-mm-diameter steel ball and a pretwisted beam that is
glued to an axially poled piezoelectric transducer see Fig. 8
for details of the geometry. By applying a sinusoidal electrical input to the transducer at a resonance frequency of the
stator assembly, the stator tip is excited to vibrate with sufficient amplitude in the form described by Eq. 1 to induce
rapid rotor rotation. Since the systems resonance frequencies are fixed by its geometry and material properties, the
motor has two adjustable experimental parameters: the amplitude of the input voltage, altered via the use of a signal
generator; and the rotor preload, altered by adjusting the distance between the steel ball and a magnet placed above it.

FIG. 8. Geometric details of our SWUM not to scale. Note that

this SWUM Ref. 8 is a simple prototype with preload provided
by gravity.

046201-7

PHYSICAL REVIEW E 80, 046201 2009

LIU, FRIEND, AND YEO

ae
mm

ag
mm

kHz

g
m / s2

0.1

0.5

108

30

1.25

6.3

70

10

A. Model parameters

In order to make direct comparisons between the SWUM

and our model, we need to know the correspondence between the experimental conditions and the model parameters.
A total of nine input parameters ag, ae, , g, , , A, ,
are required to perform dimensional simulations, however,
only four ag, ae, , g are accurately known: ae and ag
depend on the design geometry of the SWUM, is set by the
signal generator at a resonance frequency, and g can be adjusted and inferred from the measured weight of the motor
setup. We thus compare the motor and the model indirectly
by estimating the remaining parameters , , A, , and
using them as default values about which we study the sensitivity of the system to variations in the five parameters.
The default values of the parameters used in our simulations are shown in Table III; justifications for the values we
selected for , , A, , and are detailed below. We were
unable to measure A, , and during the motors operation
because the rotor blocks the access our instrumentation requires to the stator tip; the stator motion is thus estimated by
measuring its vibration without the rotor, which shows that
A 107 m and 30 rad/ s. Assuming that the presence
of the rotor decreases A more than , we chose A
= 108 m as the default parameter while keeping
= 30 rad/ s. The phase lag is assumed to be close to that of
the free vibration of the stator, and thus we chose = 0.
We estimated the coefficient of restitution by dropping a
20 mm steel ball from a height of 4.5 cm onto an aluminum
beam sitting on a laboratory bench in a configuration similar
to the motor; the resulting rebound height of 0.9 0.5 cm
implies = 0.4 0.1 at an impact speed of 0.95 m/s. The
coefficient of friction is estimated by measuring the rate at
which the rotor decelerates brake when power input to the
motor is switched off 18; in the absence of stator vibration,
Eq. 2b implies that

ma2g brake
.
a eN

44.3
t
7.77
0

4
6
time
s

FIG. 9. Color online Determining the coefficient of friction

between the rotor and the stator. A curve of the form = braket
tc is fitted to the deceleration profile when the motor is switched
off; is then inferred from brake through Eq. 17. The right hand
figure shows the measured at various preloads; in accordance
with Coulomb friction, is relatively independent of preload and
has an average of 0.10 and standard deviation of 0.06.
B. Comparison with experimental results

In Fig. 10, transient speed curves predicted by our model

at the default parameter values with A varied across its expected range, are shown next to measurements from three
separate trials of our SWUM. The two sets of curves are
qualitatively very similar: not only does our model predict
the speed transients curves to be of the observed form
ss1 et/tc, it also accounts for the presence of what appears to be a high frequency, random oscillation that existing
SWUM models have been unable to explain. In our model,
the oscillation originates from the vertical rotor motion that
is undergoing long period chattering possibly with transient
chaos, and is strongly affected by A: when A is small, collisions are more frequent and occur at lower speeds, leading

a

b

c

17

The deceleration profiles were obtained for a range of preloads see Fig. 9 giving us an average of 0.1 with a
standard deviation of 0.06. Although reference values of
between steel and aluminum under dry sliding conditions is
about three times the results of our measurements in the
0.30.4 range 19, the difference may be explained by the
fact that the stator and rotor are under line contact, which
reduces the true contact area and correspondingly the coefficient of friction 20.

0
80 160 240 320
preload
mNm

d

30

rotor speed
rads

rad/s

0.1

20
10
0
0
30

23.9
1et0.854
2

6

e

20
10
0
0
30

20.8
1et0.743
2

6

f 

20
10
0
0

21.9
1et0.755
2
4
time
s

rotor speed
rads

A
m

0.2

rotor speed
rads

rad

rotor speed
rads

rotor speed
rads

0.3

30

20
0.1
10

Known

rotor speed
rads

Unknown

rotor speed
rads

TABLE III. Default values for the dimensional parameters

30

18.3
1et0.788

20
10
0
0
30

16.
1et0.924

20
10
0
0
30

16.2
1et0.804

20
10
0
0

2
4
time
s

FIG. 10. Color online Transient speed curves each fitted with
an equation of the form ss1 et/tc. ac Simulation results at
full preload for the following values of A: a 109 m, b 108 m,
and c 107 m; all other parameters take on the default values in
Table III. df Measured behavior of our SWUM at full preload
and the same voltage input 0.55 V peak-to-peak on three separate
trials.

046201-8

PHYSICAL REVIEW E 80, 046201 2009

ROTATING BOUNCING DISKS, TOSSING PIZZA DOUGH,

30
20

10

0
0

0.03

steady state speed

30

10
0
0

0.02
0.01

80 160 240 320

preload
mN

c

20

stall torque
mNm

b
0.04

80 160 240 320

preload
mN

0

0 80 160 240 320
preload
mN

d
0.04

stall torque
mNm

steady state speed

a

0.03

0.01






0
0 80 160 240 320
preload
mN
0.02

FIG. 11. Color online Effect of preload on steady-state speed

and stall torque at two of the motors operating frequencies: ab
= 70 kHz with clockwise rotor rotation, and cd
= 184 kHz with anticlockwise rotor rotation. The measured results
are marked by crosses while the simulation results are marked
by dots and joined to guide the eye. Points joined by solid,
dashed, and dotted lines are, respectively, simulations performed at
A = 109, 108, and 107 m.

to smoother transient speed curves; when A is large, infrequent high speed collisions results in large oscillations in
rotor speed.
Quantitative comparisons are avoided since accurate values of five parameters are unknown; in fact our sensitivity
study in the next section shows that the five parameters may
be adjusted for a perfect match. We merely comment that
reasonable values were chosen for our estimates and the resulting predictions are of the same order of magnitude as the
measurements.
We thus focus on qualitative comparisons, and have found
good agreement between the observed and predicted trends
in steady-state speed and stall torque when preload is increased. In our model, the effects of preload are apparent in
two ways: explicitly, through the parameter g, and implicitly,
by modifying the stator motion. Except for the general expectation that the stator vibration amplitude will decrease
when g is increased, A and and are unknown functions
of g. We conjecture that A is more significantly affected by g
than and , and thus in determining the effect of preload
in our model, we varied g between 0.1 and 10 m / s2 at three
different axial vibration amplitudes, starting with A
= 107 m at low preloads and reducing to 108 and 109 m
in the presence of preload the default values were used for
and .
Our experimental study 8 of SWUMs showed that for a
large range of preloads between 80 and 280 mNm, the stall
torque is proportional to preload, while the steady-state
speed is relatively independent of preload. We can see from
Fig. 11, where we overlay the simulation results with the
experimental data, that similar trends are predicted by our
bouncing-disk model when preload is in the 80280 mNm
range. Note also that the response of our model at the default
frequency = 70 kHz and the second operating frequency

= 184 kHz are both consistent with experimental observation, adding further support that our model has correctly captured the stator-rotor momentum transfer process.
However, there are some features of the SWUM that our
simulations fail to reproduce: for the = 70 kHz case, there
is a sharp fall in steady-state speed at low and high preloads,
and the plateau in the stall torque when preload is greater
than 200 mNm; for the = 184 kHz case, there is a sharp
rise in steady-state speed at high preloads. The discrepancies
at high preload may be caused by the following: first,
would be reduced as the preload is increased due to the suppression of stator axial vibration and therefore torsional vibration through the coupling mechanism within the twisted
stator structurethis could explain the drop in steady-state
speed at high preload when = 70 kHz; and second, a qualitative change in the motors dynamics can occur at high preload if A is reduced such that 1 / , in which case the
stator acceleration would be insufficient for the stator-rotor
separation to occur, and the rotor would no longer be
bouncing. The differences at low preload may be caused
by the assumption that g is constant in our model, which may
be valid when the axial motion of the rotor is small relative
to its distance to the magnet used to adjust the preload, but
fails as the magnet is drawn close to the rotor to achieve a
low preload.
C. Predicted behavior of SWUMs

When the transient speed curves are of the form ss1

et/tc, we can compare simulated and observed results by
their steady-state speed ss and the stall torque ss / tc. Note,
however, that this transient response occurs in our bouncingdisk model when the vertical motion of the rotor undergoes
long period chattering that traces the shadow of a chaotic
strange attractor; for different parameters, the bouncing-disk
system may undergo periodic or low period chattering orbits
that give rise to qualitatively very different speed trajectories. Figure 12 shows an example of the speed trajectory that
results from a low period chattering orbit: the periodic nature
of the of orbit causes the angular momentum transfer to occur in a regular pattern, resulting in the linear rise in rotational speed. The angular impulse is limited by the relative
angular speed of collision; thus when the rotor reaches the
rotary speed of the stator at impact, there is an abrupt transition from the apparent constant acceleration to steady-state
rotation. Note also that the speed trajectory appears to form
multiple lines as the rotor cycles through collisions with different impact speeds.
So far, piecewise linear transient speed curves of this form
have not yet been observed in SWUMs. We suspect that real
motors operate in a region of the parameter space where low
period chattering orbits are uncommon, and that they may be
reproduced experimentally under the right conditions. This
leads us to the results of our sensitivity study of the five
parameters with estimated default values, where we found
that the coefficient of restitution appears to have a strong
effect on the presence of low period chattering orbits.
We varied each of the parameters , , A, , and
individually over their expected range while keeping all

046201-9

PHYSICAL REVIEW E 80, 046201 2009

LIU, FRIEND, AND YEO

b

20
10
0
0

21.5
1
e
30

60
time
s

150

0.5

90

d

0
2

0.5

0
w A
mms

rotor speed
rads

50

100

t18.3

c

1 + w
H
A

10

2
10
1
0
0

We have seen from Fig. 10 that A has a strong effect on

the oscillation of the speed-time curve, thus it may seem
curious that the steady-state speed and stall torque are not
affected significantly by A even when it is varied by two
orders of magnitude. A possible explanation for this is that
caused by larger collision
the increased angular impulse H
speed wA is offset by the longer interval between collisions
T

100

30

12 18
time
s

24

0.5

0
2

0.5

FIG. 12. Color online Comparison of the simulated transient

speed curves when the rotors vertical displacement undergoes long
period and low period chattering orbits. The default model parameters in Table III are used except for: ab A = 107 m and g
= 0.5 m / s2 in the long period chattering case, and de A
= 108 m and g = 1 m / s2 in the low period chattering case.

other parameters at their default values. The predicted effect

of each parameter on the steady-state speed and stall torque
are shown in Fig. 13. Most of our simulations yielded transient speed curves of the form ss1 et/tc, however, when
0.5, low period chattering becomes the prevalent behavior of our model. With our estimated to be 0.4 0.1 at 0.95
m/s and theories predicting to rise as collision speed is
decreased 21, this may explain why low period chattering
orbits have not been observed experimentally and suggests
we should reduce in attempting to reproduce these orbits in
SWUMs.
While the chief purpose of our sensitivity study is to determine the likely effect of errors in our estimated parameters, it also provides predictions that may be used for future
comparisons with SWUMs and suggests which parameters
are important for improving the motor. From Fig. 13, we can
see that an error in some of the parameters will have a
greater significance than the others; for example, the steadystate speed is most sensitive to , , and , while the stall
torque is most sensitive to , , and . Note that and
appear to be the only two parameters that significantly affect
both steady-state speed and stall torque; their effects are,
however, limited to reducing steady-state speed and stall
torque from the maximum values that are set, respectively,
by and .
steady state speed 

T wA/g

/Ttorque g1 + .
implies H

20

and

Although the above argument also seems to explain the independence of stall torque to , and the linear relationship
between and the stall torque, it should be noted that this
simplistic argument neglects the fact that collisions occur
over a range of phases and speeds, and ultimately fails to
correctly predict the effect of .
Consideration of how the collisions of the orbit are distributed in the state space phase space is needed for explaining the effects of and . The results for corresponds
to a cosine curve because the collisions are more concentrated at 0, thus the momentum transfer occurs predominantly when the angular velocity of the stator is
= cos. The collisions are less concentrated at = 0
when is increased, thus leading to the decrease in both
steady-state speed and stall torque.

VI. CONCLUSIONS

Pizza tossing and SWUMs share a common mechanism

for converting reciprocating input into continuous rotary motion, and the physical behavior of both may be represented
by the convenient adaptation of the bouncing ball model
2,3 to include rotation. Key features of the motion transfer
process such as impact, separation and stick-slip frictional
torque are captured by our bouncing-disk model, giving us a
better understanding of SWUMs and pizza tossing. While
there are many factors influencing the choice of tossing motion that we did not consider in detail, our bouncing-disk
model shows that aspects of the tossing motion employed in
the two basic pizza tossing modes do optimize certain performance measures: the linear trajectory used in single tosses
maximizes rotation speed and efficiency, while the semielliptical motion used in multiple tosses maximizes the ease of
maintaining the dough in the period 1 orbit that is required

a

b

1
0.18
0.8










0.12
0.6




0.4


0.06


0.2


0
0
0
0.2
0.4

1
0.06
1
0.06





0.8

0.03


0.5











0.6 periodic



0.04





0
0











0.4 orbit
0.02





0.5
0.03



0.2






0
0
1
0.06
0.
0.5
1.
0

rad

c

d

e

1
0.06
0.8









0.04
0.6





0.4
0.02
0.2
0
0
109 108 107
A
m

1.
0.06
0.8













0.6






0.04
0.4
0.02
0.2
0.
0
3
30
300

rads

stall torque
mNm

30
w A
mms

rotor speed
rads

a

FIG. 13. Color online Effect of varying a , b , c , d A, and e about the default parameter values. In each plot, the
steady-state speed predictions are denoted by crosses with scales on the left hand frame, while the stall torque predictions are denoted
by dots with scales on the right hand frame. Note that the rotor speed is nondimensionalized with respect to .
046201-10

PHYSICAL REVIEW E 80, 046201 2009

Estimated diameter
d d0

ROTATING BOUNCING DISKS, TOSSING PIZZA DOUGH,

for maximum steady-state speed. We applied our bouncingdisk model to our SWUM, using estimated parameters, and
found that it reproduces the oscillations that have been observed in the motors transient speed curves and have not
been explained by existing models. Additionally, the predicted effect of preload on the steady-state speed and stall
torque agrees with experimental observation.
APPENDIX: NEGLECTING DOUGH DEFORMATION

In the bouncing-disk model, the effects of dough deformation during the impact and contact phase of pizza tossing
may be modeled with an appropriate choice of friction coefficient and coefficient of restitution . What is not modeled
is the change in dough geometry, which causes changes to
the moment of inertia and aerodynamic forces. However, as
we will show below with data extracted from videos of pizza
tossing, the change per toss in dough diameter is relatively
small, thus the essential character of the process is the same
as the bouncing-disk system.
A frame by frame diameter estimate of the dough in the
multiple-toss video 11 starting from the end of the second

1
2
3
4
5
6
7
8
9
10

11

P. Pieraski, Phys. Rev. A 37, 1782 1988.

N. B. Tufillaro, Phys. Rev. E 50, 4509 1994.
J. M. Luck and A. Mehta, Phys. Rev. E 48, 3988 1993.
J.-C. Gminard and C. Laroche, Phys. Rev. E 68, 031305
2003.
K. Szymanski and Y. Labaye, Phys. Rev. E 59, 2863 1999.
N. B. Tufillaro and A. M. Albano, Am. J. Phys. 54, 939
1986.
K. Nakamura, M. Kurosawa, and S. Ueha, IEEE Trans. Ultrason. Ferroelectr. Freq. Control 40, 395 1993.
D. Wajchman, K.-C. Liu, J. Friend, and L. Yeo, IEEE Trans.
Ultrason. Ferroelectr. Freq. Control 55, 832 2008.
B. Watson, J. Friend, and L. Yeo, Journal of Micromechanics
and Microengineering 19, 022001 2009.
Video footage of single pizza tosses performed by Noah Elhage, captured on 3 October, 2007, camera operated by KuangChen Liu.
Online videos of multiple-toss, accessed 29 April, 2008 http://
www.throwdough.com/trick_basictoss.htm
and
http://
www.youtube.com/watch?vfAW0aDWaiSI.

1.5
1.0
0.5
0.0

5 6 7
Time
s

10

11 12

FIG. 14. Color online Frame by frame estimate of dough diameter from the video of the mutliple-toss Ref. 11. The equation
for the least-squares best fit is d / d0 = 1 + 0.0499t 2.365. The start
of the nth toss is marked with an arrow.

toss at t = 2.365 s is shown in Fig. 14; the diameter was

measured in arbitrary units, fitted with a least-squares best
fit, and finally the diameter was normalized with respect to
the best fit diameter d0 at t = 2.365 s. The start of each toss is
marked with an arrow. From n = 3 to n = 12 the diameter grew
by 30%, which is an average of 3% per toss.

12 K. Nakamura, M. Kurosawa, and S. Ueha, IEEE Trans. Ultrason. Ferroelectr. Freq. Control 38, 188 1991.
13 J. Guo, S. Gong, H. Guo, X. Liu, and K. Ji, IEEE Trans.
Ultrason. Ferroelectr. Freq. Control 51, 387 2004.
14 J. Tsujino, Smart Mater. Struct. 7, 345 1998.
15 K.-C. Liu, J. Friend, and L. Yeo, EPL 85, 60002 2009.
16 H. L. Nusse and J. A. Yorke, Dynamics: Numerical Explorations Springer-Verlag, Berlin, 1994.
17 J. P. Laraudogoitia, in The Stanford Encyclopedia of Philosophy, edited by E. N. Zalta Stanford University, Palo Alto, CA,
2008.
18 K. Nakamura, M. Kurosawa, H. Kurebayashi, and S. Ueha,
IEEE Trans. Ultrason. Ferroelectr. Freq. Control 38, 481
1991.
19 CRC Handbook of Engineering Tables, edited by R. C. Dorf
CRC Press, Boca Raton, FL, 2004.
20 shppE.-S. Yoon, R. A. Singh, H.-J. Oh, and H. Kong, Wear
259, 1424 2005.
21 G. Kuwabara and K. Kono, Jpn. J. Appl. Phys. 26, 1230
1987.

046201-11